A Modality-Mixing Fusion Network for Accurate Land Use/Land Cover Segmentation

Accurateland use and land cover (LULC) segmentation from remote sensing images is consistently challenging, especially when integrating different data sources such as optical and synthetic aperture radar (SAR) images. Although earlier multimodal approaches have shown promising results, many do not sufficiently utilize fine-grained cross-modal relationships or adaptively manage the spatial and semantic variability found in complex and diverse LULC categories. To address these challenges, we introduce MMixF-Net (Modality-Mixing Fusion Network), a novel deep convolutional architecture specifically designed for multimodal LULC segmentation. Our proposed model features a multistage fusion process with a modality-mixing tokenizer for early patch-level alignment, a cross-modality attention mechanism, and a gating module for dynamic feature interaction. An adaptive decoder with attention-guided skip connections is also incorporated to improve boundary accuracy and class distinction. Evaluations on the WHU-OPT-SAR dataset indicate that our model achieves an overall accuracy (OA) of 89.4%, a mean Intersection over Union (mIoU) of 52.8%, and a Kappa coefficient of 75.7%, surpassing several recent benchmarks. These results highlight the effectiveness of MMixF-Net to effectively integrate multisource remote sensing data and highlight its significance in advancing machine learning-based geospatial analysis and environmental mapping.